14 DAY TRIAL // In what may very well be one of the biggest overreactions to miscommunication ever, buyers of the high-end GeForce GTX 970 graphics card have been on NVIDIA's case for a few days, to the point where the company is actually suggesting that they return the board.
That is not quite how they put it, but an NVIDIA employee by the name of Peter practically spelled it out in one of the latest posts on the GeForce forums.
Why has it come to this? It's all because of the memory allocation problems discovered in the graphics card. We have documented the saga quite closely.
How things snowballed
The drama started when users began to see visible drops in game performance when the GTX 970 tried to access more than 3.5 GB of the 4 GB VRAM.
One thing led to another and it was discovered that 0.5 GB were allocated separately and differently from the other 3.5 GB, due to how the GPU's SMMs were set up.
Later, a diagram was published featuring the GPU's composition and memory allocation, but with it came the discovery that NVIDIA had released wrong GTX 970 specs back in September 2014.
While it was all attributed to a misunderstanding, the confusion occurred between the engineering team and the PR team, so NVIDIA is taking responsibility.
NVIDIA is preparing a new driver
The driver will be able to “tune what is allocated where in memory” and will further improve performance, or at least prevent some of the erratic drops and highs that occur when the card tries to access the 0.5 GB.
However, some consumers may want to replace the board regardless, since they were basically provided with false information. NVIDIA's employee said that either a refund or exchange would be easy enough to do via authorized representatives.
All par for the course for a mishap like this we suppose, if not for one thing.
The irony
Strange memory allocation or not, the GeForce GTX 970 is, at the end of the day, still the best graphics card you'll find at its price point. Benchmark tests and game performance show that, regardless of the fact that the number of SMMs is lower than initially stated, or that the cache is of 1.2 MB instead of 2 MB.